While there’s no magic eight ball for medicine, Yale researchers have managed to predict how likely someone is to develop skin cancer.

Their approach, published in the journal Scientific Reports on Jan. 26, uses simple metrics such as age gender and race, to predict how likely someone is to develop nonmelanoma skin cancer. The new model may help shape the future of cancer diagnosis.

“A lot of people aren’t at the dermatologist’s office, and yet they may be at high risk for skin cancer,” said Christine Ko, a professor and dermatologist at the School of Medicine who is one of the authors of the study.

She said that most people who regularly visit a dermatologist for checkups have a family history of skin cancer, have gotten several bad sunburns or have skin lesions, even though other demographics may be at risk as well. And while many are conscious of the risk of developing melanoma, other types of skin cancer may go unrecognized until more noticeable symptoms appear.

Using a type of artificial intelligence called a neural network, the researchers fed massive amounts of data into their program to train it to link specific inputs with risk for skin cancer.

“The neural network has the ability to decipher a hidden correlation that is impossible for human beings,” said corresponding author Jun Deng, a professor at the School of Medicine. “Our model is actually not to say that this person is going to get [skin cancer]; rather, it is proactive and preventative.”

In this way, their approach represents a means to refer high-risk patients to a dermatologist’s office for a more thorough screening. Catching people likely to develop skin cancer before the manifestation of the disease is one of the main goals of this technology — which Deng calls “early cancer detection and prevention.”

“The utility of this kind of model is for more preventative care [and] to figure out a risk factor that we don’t know about,” Ko said.

To train and validate the neural network, the researchers extracted parameters from the National Health Interview Survey 1997-2015 big data sets with 463,080 respondents. The data sets were taken from the Centers for Disease Control and Prevention. About 87,500 people participate in the survey each year, providing health information to the U.S. Census Bureau. Deng and his team downloaded the freely available data set and identified their parameters from there.

Their model used 13 simple inputs to predict skin cancer likelihood: gender, age, BMI, diabetic status, smoking status, emphysema status, asthma status, race, ethnicity, blood pressure, heart health, exercise habits and history of stroke.

According to researchers interviewed, the neural network’s predictions were especially impressive since their data inputs did not include UV exposure or family history of skin cancer — the two most common factors dermatologists use to determine skin cancer risk.

Using artificial neural networks may raise new possibilities for predicting other types of cancer.

In addition to working on skin cancer, Deng studies prostate, lung, pancreatic and colorectal cancers, and he is working on applying the network technique to predict risk for developing these cancers.

Deng added that the kind of technology used in the study can even be developed into a mobile app.

“The patient can actually do it by themselves, at home, without a physician,” Deng said.

Approximately 9,500 people are diagnosed each day with skin cancer in the U.S., according to the American Academy of Dermatology.

Jessica Pevner | jessica.pevner@yale.edu